Debugging Executions

Use this skill when debugging workflow execution failures or successful runs with wrong or empty values.

When the user reports it still fails

Re-run the failing path with executions(action="run") (or verify-built-workflow) and inspect the real result before responding. Do not restate that the workflow is "fixed", "verified", or "working", and do not attribute the reported failure to a test-harness artifact, stale state, or "it works in production" without a re-run against the failing path. Treat live signals as real: an execution error, partial coverage (nodesNotReached), an empty node, or a missing node is a real defect to investigate, not something to explain away. If you genuinely cannot re-run the failing path, say so plainly and name what is unconfirmed instead of repeating a success claim.

Draft versus live

A published workflow runs the version that was published, not the draft you edit. Your save creates a draft, and the draft is not live until somebody publishes it. So a fix to a live workflow changes nothing in production on its own.

Two consequences when the user reports a live workflow failing:

• Read the version each execution ran. executions(action="list", workflowId) returns workflowVersionId on every row, plus workflow.activeVersionId (the published version) and workflow.draftVersionId. Only a row whose workflowVersionId equals workflow.activeVersionId ran the published code. A run of a draft proves nothing about production. A workflow.draftVersionId different from workflow.activeVersionId means the latest changes, including any fix you just made, are not live.
• Name the version when you invite a retest. Say whether the user tests the draft or the published version. "Try it again" after a draft-only fix sends the user to the broken published version, and the fix looks like it failed.

Do not report a fix as live until the published version is the fixed one. Ask whether to publish it instead. After a publish, a new live run is what confirms the fix — an earlier execution ran the old version.

Testing event-triggered workflows

Use executions(action="run") with inputData matching the trigger's output shape — do not rebuild the workflow with a Manual Trigger. For trigger inputData shapes, read ${N8N_WORKSPACE_DIR}/knowledge-base/reference/trigger-input-data-shapes.md when a sandbox workspace is available.

Failed execution

executions(action="debug") already includes failedNode.resolvedParameters — start there. That bundle has parameters (raw, with expressions intact), resolved (substituted), failedExpressions (those that threw), and emptyResolutions (those that resolved to null/undefined/"" silently). The offending expression is usually visible without a follow-up call. Entries in either list tagged with reason: "unreconstructable-context" are NOT real bugs — they reference variables we don't reconstruct in replay ($vars, $secrets, $response, $request, $pageCount, $ai). The value existed at execution time; we just don't have it here.

Confirming a fix on the node that failed

executions(action="debug") tells you what the node received. It does not tell you whether your fix works. To learn that, run the node itself:

executions(action="run-step", workflowId, nodeName, reuseExecutionId=<the failed execution>)

reuseExecutionId replays the data the node really received and re-runs only that node, so the fix meets the same input that broke it. This is the right first move whenever the user is debugging a read node that already failed a real execution: the node ran for real once already, and a mock-only check is what sends the user back for a second session.

The action refuses a replay that would re-run a node above the target: an unpinned one the execution never reached, one that failed there, or a loop that did not finish. Pick another execution, or tell the user which node would run again.

Decide whether the node is safe to run first

A step run is a real run. The node uses the user's real credentials and reaches the user's real systems, on the user's real data. Check what the node does before you reach for run-step:

• Safe to run. A read (get, getAll, search, list, download, a GET HTTP Request), or a transform that touches nothing outside the workflow (Set, IF, Filter, Code without network or filesystem access). Run these.
• Do not run for real. A write (create, update, upsert, delete, send, append, a non-GET HTTP Request). Running one sends the message, charges the card, or deletes the row — again, and for real. The user asked you to debug the node, not to perform its effect.
• Unsure? Treat it as a write. Losing a debugging shortcut is recoverable; an un-asked write to the user's data is not.

For a write node, debug without running it: read the failed execution with debug, inspect the resolved parameters with get-resolved-node-parameters, and explain the fix. That is usually enough, because a write node's failures are nearly always in its input or its parameters, both of which you can see without sending anything.

If you genuinely cannot resolve it without a real run, say plainly what the node will do to the user's data, and let the user choose. The approval prompt alone is not consent: the user sees a node name, not "this posts to your #general channel".

"It already ran anyway" is not a reason. It holds only for a node that errored outright and changed nothing. A node that partly succeeded before it failed — a send that delivered some messages and then hit a rate limit — will deliver them again.

Studying a node on its own with mockInput

mockInput runs the node on items you supply and skips everything above it. This is a good way to study one node by itself, and a normal thing to do while debugging:

• probe an edge case the workflow rarely produces — an empty list, a missing field, a zero or negative amount;
• hold the input still when the upstream data changes between runs, so two attempts are comparable;
• separate "this node is wrong" from "this node gets the wrong input".

Reach for it whenever the question is about the node. Use reuseExecutionId or a chain run when the question is about the workflow.

Keep the claim at the level of the evidence. A mocked run shows the node handles the input you gave it; it shows nothing about what the chain really produces. The result carries inputMode: "mocked" and a mockedNodeNames list — report the node's behaviour, not the workflow's.

Mocked input does not make a write node safe. The node still runs for real against the user's systems; only its input is invented, which makes the effect less predictable, not more.

Running a tool

A tool never runs on its own. n8n runs it through the node that owns it — usually the Agent. A step run on a tool therefore behaves like a step run on that Agent:

• mockInput feeds the Agent's input, not the tool's arguments.
• reuseExecutionId replays the nodes above the Agent.
• A chain run (neither option) runs every node above the Agent for real. Give one of the two options when a node up there writes.
• ranThroughNodeNames lists every node that can run the tool, not the one that ran it. A tool on several agents lists them all: the engine picks one, and it does not report which. Do not name a single Agent to the user when this field holds more than one.
• A tool on two Agents where one runs above the other is refused. Run the upper Agent, then read the tool with executions(action="get-node-output").

The tool's own arguments come from toolArguments — the values the agent would normally decide:

executions(action="run-step", workflowId, nodeName="Search Tickets Tool",
           reuseExecutionId=<the failed execution>,
           toolArguments={"query": "login fails", "status": "open"})

The example targets a read tool on purpose. The write rule above holds here too, and a tool hides the write behind a friendly name: a step run on a "Create Ticket" tool creates the ticket again, and toolArguments does not change that. Only run one when the user has accepted a second write.

Use the argument names from the node's $fromAI calls, which workflows(action="get-as-code") shows. Pass a plain string instead for a tool that takes one free-text input (Wikipedia, Code Tool, a vector store used as a tool).

A tool that declares $fromAI arguments is refused without them: it would otherwise fail for a reason that has nothing to do with the user's problem, and you would report that as the defect. A node that holds several tools is refused outright, because nothing here can name one of its tools the way the agent does: the "MCP Client Tool" node, and every node the MCP registry added, whose type is @n8n/mcp-registry.<server slug>. Run the Agent for those, and read the node's output from that execution.

A sub-node that is not a tool — a model, memory, embeddings — cannot be run this way at all. Run the Agent, and read the sub-node with executions(action="get-node-output") on that execution: n8n records every call a sub-node made while the Agent ran.

Successful execution with wrong or empty value

When debug doesn't apply because nothing errored, call executions(action="get-resolved-node-parameters", executionId, nodeName) on the node whose output looks off — do this unprompted, don't ask the user for permission first. It's a cheap read-only inspection and the only reliable way to confirm whether an empty value came from an expression silently resolving to nullish. Check emptyResolutions first; most "this parameter is empty" cases are expressions resolving to null/undefined/"", not thrown errors.